a N Te ietes Lathan vies: . Foundations ef Newel N tier ks Le the biokayat Newton Ly Neurad Nel Guvies ly P tan a Ss male Larger Le mo legen Pore Ram Ls Traicing reveal netoork bs Ah bvahion Aonchinns > Unca, > Siqmosd > Tan =H tanh —> Soltmos > Re chifed Linco Cretu) Loss fun han > Notctisn, Regrecsron Man abilute em lou CMAe) > Mon Sapawed foq Er (mile) > mean abte kite Per lenfane lass CMapey » * Kotzan by ype par emetey 4 (tied tess = leavnary yok; be logtiaic oss > GartsipH ANCUNAS Networks => fh Neral netiyorty Xo o typ, & Compaley PP Fran yhot 2 dened {0 Wor)s Glee tie fuman brat, ) Smale Ports CUeel Newrong ah oO WE So mode UP ot, mon » WHICH ape Connected 10 @ thew. breuh Rarvon, to oth ners, This , baw the pouvet nbowrk to fawn Brom the dat, ¢ » make fre ArctBp ns } tp “She (ane Ctton'y bho nodes thane weal, what, a the lonnec Bon me The most Well toe Sie Plast fo cardia fae est dhe feed word valent d thos rpt bye ene) Adder Joutn go ie ) ) otermi ng the Strexgt ) | ofp Monten, Oat daywe Ne 3 3 (4) The Rinio pal Neuvon, Dendrite Axon €rminal Node of Ranvier Cell body Sg Mp Sehwann cal Ss Myelin sheath leus => Neurons 2% DL were Lnspived by nauaons En the heenny tain. Were ne G@nfral Parte] the rowvon Conte nine) “The rechuny Denote, ‘| Berend, Like Gbevs extending fom cul body. Pordrite eevee $°qrabs fem Sorterrinding neuron x 3 A Lixo | wed ffbet thet Prams b checto ntheseak ToPulre Baqeel do other noone, CAL pouve hay one Axon Kx loin fe MOVE Sh attedww a> PersleP tyon te XQ malig | Bacle Prpogebon ola Ss) —s ——— ~ eb A pr ceptwn ds one of Ahe Serp lest (ape ey pawveel phous : aa” By uted fe inany clws|g calion « => A pera phon tr a Sraledoy narnal padevorls Cond’siling of jnged foes, weights ard activation (enter A> at telres 8n multiph input patios Cate ptt vali “2 an nig acd Lomi ny Ahem wit Corves Penal ee > these * ais dnpals ane Sumered v spard the vesltric Pored, Howugh the [rctaveion function to Por Prodiste ox Of) wanking | @ 1O8s adtivation (eye 1 7 T° xO \ fertion & to beo YO" Ouptyee LCs Tx OO 40) Sr ak ae 2 © NO we Us the Loge ef Av getei-c ve dave focus wert bho Offic gto %rO vO | »6 San eta vit yO \~ 50 ¥3 90 vl oO “ betonana 49 AND ©) 29 WOT. = ve if ee pod : wl) VeOO — SS ee EE el, © tab 0 in Place 0f # Hei ® 2 B1S $0 49 ech re O—> Ack & Pat 1 dm pote of x andffzo) tn @ oO Slo +) fos -0-5 = TO put x20 end ye) oD i Saby Wty s or So @ pab x=) and Yet fn ey | = AS the 2.08 = 4 Hs, tf baci futin bx OST oshich mond tho bo dake oD @Qs-os ec « t Lo Gotatee o aber @O s-os uv tLe Stake d @ Bae oe txo Sottke | Wofes- Dusk ng teaiaing, the Perception adiust 25 WE {ne and las {recporeree Vat influence spur beibiondy “badadt om Tae jpit ARR ad ty desived Of P athe pe a> 40 mins mige We Predi chisn erocty foting Tne wotgnts abd bal wig an Opa aon off thon Li Re"> nol Baye Aap on Ep lion 6 yfudspypn 3% here muthiple lowers “f, nouns ( nods ) ounrarae| an a feed fonwrd rae Stwwluve > An ME P Comists oan input Larger ) One GO more hidolon , , ) ’ ’ eee ard on oullpid Jorer ’ y , , > Wording ropa Sayer ; at , lanier ser thus ame felrmtith Joyors Soon tne Pe and octpt ortecss Ipckivesn —pre ton pst Eau nun Ao Wadden harper ret input rom all nerves on the Pretirg > Large, applies a wetghtid bam and an achivetin fete fares the Yerutl +0 dwn reuwrs zn ‘the ast aps =o titpul fovferd Tease toge proclucos the yoo coll of the weaons eerie -(a) “1 OH) war | i = Se Dobpal 3 Fs) x i =) Co 5 , ’ , , , 'BM Now Cohuteling we Pb Ikon log nocls too qere wo olaleho = WR andwot aan Ou ; Use wis (#1 tun Gayl = 2a) # CY lo) =~ 2 —» Fo Sotakey =I Wry axl Wr4 aare Ovy = |Vasey Gel + Wry (x23) Vue ICD # 4 Ces -5t29 talee) a ee a = o Vee Va( wir) yy (owe tye X1 CD iF - = DIE) eo agen Paley | a, &) Training neural patcoavk 2 = —= — — Training, rawred relivorle crar|ves teothing a +o make ocuwedt predtelions (rouauilj vations boy | ad3us Brg 20 Pours meters ( weighty Gots) bared on fot date ancl cured output a lw ace y Crodiew anck oM dataset. Ths frome b> Cre protons ar ) y ) Trudlve tasks Like Cleaning the dota, nav ge ©) Scaling Feats ancl Splitting tre tock sno frrertning vali delion ant batseb. % Soi bak gibin ; : by ag Dniiolize Phe wtyny @ béader © jhe pew . polionde : Ly Feed fie’ tmning examply forererd ~ro.ong b “Hie pabicork 4 Com pufe predic Fors 6D outputs. Tus Rawal ves mw khip ag The inpul feadiers by dhe weighs and ot phying attiv ation fen chon ah Rotts y pou Y Los Gla Hon by Compare le Prdicted outpuls wrk he attua Jobsb for the {pFateg training Bok po Calculate thy Lou /enr©) Back prpegadion 2 b Barkproprgelion 2d the Procers Of Prpey path err boc teumrd ~Hostuala the ator to ufdele toe wae aids and tka. Ly Adjuat the Cougs and betes Ln Hae dlnelion hat vecbcoy tn, As » typiallyusng op Fieyacbion 7 thm Silce Gradient De sent 6) Sfothase © validation Cratinak Tx hry ; phcivetion Eaten Fonction S: ty A Activalitow bm functions, Sy ov mothe matt al funilins ApPhad to tne OSPL ef cach neu» Se ies alien Lethe AE boli w @ determine the oak pul of New ie Whey dete seater fie neuron Shoot be edit st b thy heride whethyy cruvor Boulel be bcbpeled 7 ared on due rapt aL Ye (ove10 ne pM A Lo = poueal nolwovlss wold Vnvav yal ecffons lip an a Mucky tre WL ) oo? without actinction fontion, | only We able fo toon Stele = Activatw» > pndyoclute non Hinessily, ablowin , Rear nefivovlds do Lean Wo lox pattivs | Predtations nf? pe & make bitter aman Mn Ga >) Linson- . ——— At 49 fas; colt “the who the Aspen duct yornable thas dive fo Lode firdect vawolals [othe oatpat & Alrvetly Proper Herat tothe Papal) tte arty a Unsor actonbionfunstin, the vvelstionshcp bes fhe input ond patpul just x Steaigah Line, Oh the tpl tncvesses | Ae oulpul froveadee prportinal by aud vice vers fly, Propofol > > ) ) ) ) ) ) )ane . oat. TEN ith eee ALN “SESE A ruse poedlipg Aue peony uy Boson Od yation Leemation allows {he nfs 19 fp x vabs Gveckty | if, the ie npat ‘fpaliores igo, Sips, lod -) "ene Predicted Price abo Ga temas VO Sigmord : > The t Stgwotd gactivation fesnrition Wed Sonewrel tha ayes toring rat bf 0G) by|love hows sbworls 9 J Lb siq mroicl 2 like cs gate (capey Prat Give tee Veleey wo Oard)= LKample se Wars “4 dhe Sigmid frnitinnes 8 Gedba leas pernact Ane Crbianle Sf a panty. Th qateteespey facies ushetrey fo het people cin based on Auow> “eo! Hag aye. > The geteeeper taltes a Look at each Povson(input) and OAS nds thom a Coo (nes SOU befiveen O&¢ | PD Semtort io Super taol” Puy get asturre 0x Fe and an allored jn. 3 fuarare ns tool, fy gets Sure Cloe 0 and areturnead oni US : pale whe Siqgmaid [artis shen we Wart the neuwah rativorls do make datiaines cv) predict Probabrlites: oe 3h we ane To ping do Separate Spams & non {pany ROBE AIL ie Gite whe. Sigmaid funtiB, B) Tanh . ' > Thetanh (hyper bo hic tangent) activation fuvihion 419 "Ano thew macthamabtal {inition teal fg nawval nilworls. thet BS Are input vols bly —lard Jfg awa, ha 5 ween, a _ 7 > the tanh fvution to Midtaw fe We Ane {enction bit Sejmmalc awerd 0. Tht at oJp ave vaeles Sa -ve init (4 means af bots if , Ave vals for +ve prput \avc tre Kea tebe | Le Rejeckann / cue ptt “y os @ Hard Tanh “- ; > Sin flav fo fen; howd tarh Srmply ey pdrea havd SPs 4o the noreralized eg je A thore, PINT Han | fomrade tnko | and Any tianng faitien 2o mode. ints Ly Ot B abo a pothsmitial fraction Heol fa rowel vativorts » Par toler ly in the output Leger boy Mubhilars classi frtion ae maging ‘you fave ult ple ovtpul” nev Tr4tho Eira Aoi of your Nr kath ve presentiog a ill ctos, > The Gol tmax faction taless fae yaw uh pret has « rewrors avd CmWev i them Pht Prolechil: M1 Shasising Aok fisty dam up tod:© Rectyjed tances Crerv s w5 At As most Commonly wsed AF Dl. > Stila Simmp\e mocthemotiol fermcition prot jab veda Non- MProawity grt the Webwork =i ie ty Sh the FP fo the Switeh tye toe = Ihehnt a Sw 7 He - ap ye ib bloules the Camels OF) Loss Con ction: > Aless keuction Bra martherates fenton taf measures how well o modell predictions patth the actus! target values. —? The toss’ Lnclicedes bmw (rr My Tae model's predictions ave fem the tive vals » Ae Ao0k during typining io 49 onini onze tn teat function, do fmprove, the obih fo make axtuunat, rcs lions se TPS (E= wolaGon OD Reavy expen QD- cst odio» (- Gnitatews ) ) ) ) ) ) ) ) ,iS kD Nola trons = loss funstions one often arated using wrecthearitel natebion ftovep rdent Ihe ob rjeat ve fumeldon oct measuriy the do herente bls?» fares : oT sepenget wala) DD) Genaral rblatio, oh Me 2): Tis represent the Low funcban Ly WW tales as input trttongtvales (Yoru) and Pradealed Vales Orred 4) : ERD Reger ES TOW 3 On pL for vegreo im tesla, the mean Lqpared Grnlmse) Ass f ection Is Commo’ ubed- & (Snr Yea) I< Heres (vw Yea) = ALP shore, — ne — poumboey of, Som pls dre < Tr kegel velit fr Ihe i! SarPle fed = preditted valet for Ihe jf Gomple > Qrogine peace trying {0 Pred frowse pres Fest eet > worl! ccnp 1 BS Ofareachbouse js (ook el how fer | t+ Wows ets Sse = cost, Pe Sout wee ye 2 tesabt, Pate * 23? suo Kh “Your pred ction Xs lyri aula} 2. btowe 2 + = ve sakt,Price = THO prite. Gs doue4 ES Ze > SHB Pvc = TIP © U Sqn these dbiggevences fp malte (' Saxo pode predicls tre Bore Thay ereabl te inte wie puede) trae bowses: Loa Pra Ld ri s Hose . Orton add upell. these Squors | Eiiprente whith in dhe (mie) | fser Lanse (fue? Yord)= 2 (Bax a whe ne noarbene] Mudd iy wouse 33 Coreg)? . = ae ee EE C80 -4g)> Har 4 Smopthe Squared Aierentey Ua ¥ au te] St measures tne aoa. Dif pemeete ANd Abslute p L \ Lasse >Ypees) ee £ 32 _ sy | © Meer Sqyared nero anal Wiebe pec man nat veya : \Lmte i You) * 2 a ee 4 tie (te fu: aa ; 3) it @) ieee bids Me porte, Adore enero J Joss Lb B =e aK meohwres 7 ®| Len n, rr > Sagi > Se DL, Classi Cant iffelion dass fertiont « are Ued when lhe Feats Svalves GAlagosatng i Tnputy. 2b chifjrent Classes oy Cokegesi 4 ° eset or rrvedure atl a Ow are loss:- Trove a burch Of dalapornty , a qe wart to > Qwagind 44 { draw ae ~ ae them jan so eps, Uke Aopen patinnen Cot 4 doy 20 Piciew othe Ringe Loss onesturaa {iow "itch Atese dale por ari on Ame ant Bide ey Te Hine ( Covverdhy chown’ YB) eure wary Side ( mis chassffedd- © Th a data Point Bo wd awed bem The bine § Covveettry Clanifjed The Das pa bow: wee VUES Tod dadage’ ©: SI © Rut S eet =— | iE 4a dal point 2s on the wmng Bide ol the bn @yfo | Onn 401, the” Los Cntateades 7g Is Like Tuas Bis t Kes C3; ty vert on the signe Aide © Th ive. cand c tapas a Eo ee ge icaadadiaiiaias But iy avte JAMS = dM ave pet & gerd nm The awry ditde evisotlasd \ ' and the Los peo tp, > Dis a5 S trating, the goal Sy fo min? 3 ousting the Line, OD eteaitgbly 4 pose? ble, ize tidy overall Loss | a { So thal Zt Sepersles the ATeoondh, 4 { { { { |W @ Loris be Loss os 2 eee stoguboss = Ae 30 a Type og Clots scallion Joss frais udiedl ry Birory Cousifyatean tastes, wtheve the goo! ib t0 Predeck whethey an eral belongs 10 on eftwo Classes C4, Somer) Nit spams, tye GIVE) Frerrstds Pe _ [a ECR alti) 20-0) ore pyc ro Oh Samples Qe, e tet beable ber iterple ed = Predechid tobl for 1h Lample exemple Am OAS ne yoware Oui dig at Yan email Vossifer Yauwanl Your model to Precicl whethey an incom ing c En Spars (claw a yao ret Spam (olnssa). : for cach amsil, spur model makerror predittion, usually. So Abe foo | « probabil - Leth, Lay fire madot Preeltts thet thereds ors chante tye email xo spam ( Cla7) anc 220% Chante ab rat Spam (Chet d.cn SEE ay . 2 gah yay \s #1) HMpur 2 * om abby. = Pew Ss ~P wWL Boo arpub dot, Jo Our mrad Fe Pecrmateg cee Bet GD tune duds Povamalin dowels. w/o Frain beter 4 (ester Thre Tunaing Pairamallers ere Called Reyer Derva rcalers © Acayning Qates ® Learning vette Shows athe Step Stye at whith the mobl Parrnrnctins Cuthanbyane wpelétid chuwing run/jele a * Tet gn Reavn ng cat mot Quite the model 40 oredhyst the Optimal Solubiom, while 4o0 Low a Lerwning vate may eburk Ay Stn. en Thy > DP Ke adjusting hoo bf astep yom fake whan (pouty aretpe. Too big and you enigyt on80 inp rail AC too tll panck Zk mM tale, Wevor feat 40 the ond. - 7) ) ) ) ) , @ sparsity sg 5 Dojperpawamslens Fmplies thot only a foo WuyperParamdtins hove a Surrpacl on tee model's Perfar mont hi LE otlres moyhare minimal ermag are afi ? Pdeabihayrng Ahose Cw Heol Aa per paromsturs Con help 20 Fm Provirg the oP tmjrotion Process & Fmprve te Effjadncy eh model toning > Spars ity ea wales Con be Lebentitied Howegh tecnigps Dath od Sensi trusty anoles where how gets og Binyy Single hay pox parerntivg on ae modll's Pex foro are Syslemacdi gly expeluratad- _ x THE Ed